Detecting outlier demand in railway networks

Transport service providers, such as airlines and railways, often use revenue management to control offers and demand in mobility networks. Such systems rely on accurate demand forecasts as input for the underlying optimisation models. When changes in the market place cause demand to deviate from the expected values, revenue management controls no longer fit for the resulting outliers. Analysts can intervene if they deem the demand forecast to be inaccurate. However, existing research on judgemental forecasting highlights fallibility and bias when human decision makers are not systematically supported in such tasks. This motivates the need for automated alerts to highlight outliers and thereby support analysts.

Network effects complicate the problem of detecting outlier demand in practice. Passengers often book travel itineraries that stretch across multiple legs of a network. Thereby, they requests products that require multiple resources – seats on several legs. In the rail and long-distance coach industries, the large number of possible itineraries makes it likely for outlier demand to affect multiple legs. At the same time, outlier demand from a single itinerary may be difficult to recognise as it mixes with regular demand on individual legs. To support outlier detection in transport networks, we pre-sent a method to aggregates outlier detection across highly correlated legs. We propose to use the results of this analysis to construct a ranked alert list that can support analyst decision making. We show that by aggregating we are able to improve detection performance compared to considering each leg in isolation.

Keywords: Revenue management, Networks, Outlier detection